Network planning for distributed datacenters under probabilistic link failures

With the widespread usage of cloud service, data center networks (DCNs), serving as the underlying supporting platform becomes increasingly important. However, the traditional centralized DCNs are not so scalable and efficient to meet the rapid growth of service demands. Therefore, we propose to split the racks of a DCN into multiple groups and distribute them in a given optical network. With given transmission delay and failure probability of each link as well as DCN internal and external traffic loads, we try to realize the distributed DCN and minimize the total cost. To combine delay and failure costs together, we introduce two cost scaling factors, which can be flexibly adjusted to control their relative importance (i.e., weights) and satisfy various performance targets. Then, we formulate an integer linear program (ILP) to solve the network planning problem and carry out extensive numerical simulations to validate the performance of our scheme.